Forecasting Tech: On any given day, the European Centre for Medium-R

On any given day, the European Centre for Medium-Range Weather Forecasts runs its flagship model on a machine capable of roughly 30 petaflops of sustained compute — processing observations from more than 30 million data points ingested every six hours before a single forecast line appears on anyone's screen. That number is worth sitting with. The physics of the atmosphere doesn't care how elegant your equations are if you can't solve them fast enough to outrun the weather itself.

How the Models Actually Work

NOAA NESDIS

NOAA OPC

NOAA CPC

Numerical weather prediction (NWP) starts not with a forecast but with a problem of reconstruction. Before any model can project forward in time, it needs to know the current state of the atmosphere with as much fidelity as possible. That process — data assimilation — blends satellite retrievals, radiosonde balloon profiles, aircraft AMDAR reports, and surface observations into a coherent three-dimensional snapshot called an analysis. The ECMWF's 4D-Var assimilation scheme does this by iteratively adjusting an initial state until the model's simulated observations match the real ones across a six-hour window. It's an optimization problem solved at planetary scale, and it runs twice a day.

From that analysis, models like the GFS (NOAA's Global Forecast System) and ECMWF's IFS divide the atmosphere into a grid — currently around 9 km horizontal resolution for the IFS's high-resolution deterministic run — and step forward in time, solving the primitive equations of fluid dynamics at each grid point. Pressure, temperature, wind, moisture: all evolving together, coupled, nonlinear, sensitive to small errors that compound with every hour of lead time. By day 10, even the best deterministic forecast is largely a climatological guess dressed up in model output.

That's exactly why ensembles exist. Rather than one forecast, ensemble systems run 50 or more slightly different versions of the same model, each initialized with small perturbations to the starting conditions and, in some systems, to the physics parameterizations themselves. The spread of those runs quantifies uncertainty in a way a single deterministic forecast cannot. A tight ensemble at day 7 is a genuine signal. A spaghetti plot of 50 diverging tracks is the model telling you it doesn't know — which is itself useful information.

Where Machine Learning Enters

The ML disruption in weather forecasting is real, and it moved fast. Google DeepMind's GraphCast, published in Science in late 2023, demonstrated that a graph neural network trained on ERA5 reanalysis data — ECMWF's 83-year atmospheric reconstruction at 31 km resolution — could produce 10-day global forecasts that outperformed the ECMWF's high-resolution deterministic model on 90 percent of 1,380 verification targets. Huawei's Pangu-Weather achieved comparable results using a 3D Earth Attention Transformer architecture. Both models run a full 10-day global forecast in under a minute on a single GPU.

The reason this matters isn't just speed. Conventional NWP requires enormous operational infrastructure — supercomputers, data pipelines, software teams maintaining millions of lines of Fortran. ML models, once trained, are comparatively cheap to run, which means they could eventually democratize high-resolution forecasting for national meteorological services that currently can't afford to operate their own global models.

The honest caveat: ML models are trained on historical data and can struggle with rare, out-of-distribution events — the kinds of extreme situations where accurate forecasting matters most. They also don't yet assimilate observations in real time the way operational NWP does. The current generation is best understood as a fast, skillful post-processor of the same reanalysis data that trained it, not a replacement for the full data assimilation and physics pipeline. That gap is narrowing. Nothing major shifted on this front this week, but the research pace makes any summary feel provisional within months.

The Reanalysis Layer

One piece of the forecasting ecosystem that rarely gets public attention is reanalysis. ERA5, produced by ECMWF, is a retrospective reconstruction of global atmospheric conditions from 1940 to present, generated by running a fixed version of the IFS model in assimilation mode against all available historical observations. It's the closest thing atmospheric science has to a ground truth for the past — used to train ML models, validate climate projections, and study events that occurred before dense observational networks existed. When GraphCast says it was trained on ERA5, it means the model learned the statistical relationships embedded in eight decades of best-estimate atmospheric states. The quality of that training data is inseparable from the quality of the resulting forecasts.

Eyes on the Sky

• When a forecast shows a wide ensemble spread at day 5 or beyond, that spread is the forecast — treat the deterministic track as one scenario among many, not a commitment.
• ERA5 is publicly available through the Copernicus Climate Data Store; if you want to understand what a past weather event actually looked like at altitude, it's the right starting point.
• The ECMWF publishes its ensemble plume charts for free at ecmwf.int — the "stamp maps" showing all 51 members side by side are one of the most honest representations of forecast uncertainty available to the public.